Five Questions About Dyssonance and Its Case for Rethinking AI

Artificial intelligence concept illustration

A former Google DeepMind specialist has published an essay explaining why she left one of the world's leading AI labs and what she is building instead. Here are the key points of her argument, organised around the questions readers are most likely to ask.

According to The AI Journal, the essay, titled "The Paradox of AI Safety" and published on 18 November 2026, was written by Dr. Jennifer Prendki, founder of the AI company Dyssonance. It explains both her personal decision and the technical idea behind her new venture.

Who is behind Dyssonance?

Prendki came to AI through physics. As a child she wanted to understand the universe, and she later found in AI a comparable intellectual challenge plus an open question about what intelligence really is. At Google DeepMind she worked with leading researchers and was responsible for sensitive decisions about the data used to train Gemini. Those decisions weighed on her and sometimes kept her up at night. What stayed with her most was a growing unease about how AI was being built.

Why did she leave Google DeepMind?

She writes that she was puzzled to see many colleagues focus on improving models without asking what they were building or what would happen once people used it. She hoped that being inside would let her make a difference, although there was no guarantee her concerns would be heard. As that hope faded, she felt responsible for outcomes she could no longer meaningfully influence, so she left, even while fearing she might later watch the consequences from outside and regret it.

What problem does Dyssonance want to solve?

In Prendki's words, today's AI is good at remembering information, but remembering is not the same as understanding. Systems such as ChatGPT, Claude and Gemini still struggle to handle conflicting information, to recognise their own errors and to keep updating their understanding. She also notes that current AI is largely built around patterns in language, and that communicating an idea is not the same as understanding it. Her expectation is that the next breakthrough will come from teaching machines how to handle ambiguous situations, not from adding more data or larger memory. In other words, the bottleneck she sees is not how much a model knows but how it reasons about what it knows, especially when the pieces do not fit together.

How does the company plan to solve it?

After leaving, Prendki chose not to build another, better large language model or tools that patch the weaknesses of existing ones. Drawing on firsthand knowledge of how LLMs learn from data, she went back to first principles. Dyssonance is developing a common-sense reasoning engine, a system that develops and refines its own understanding, with interpretations that change as evidence emerges. Contradictions are treated as opportunities rather than noise. Just as scientists, detectives and entrepreneurs make progress by noticing what does not add up, AI should identify contradictions and work to resolve them, not ignore them or simply store more information.

The essay compares the difference to memorising answers for a test versus actually understanding the subject, and to reading many books without ever connecting their ideas or changing one's opinions. The goal is a system that behaves less like an archive and more like a careful thinker, one that notices when a new fact clashes with an old belief and revises its view accordingly.

What is the bigger picture?

The long-term vision, which Prendki calls engineering intuition, is a shared underlying foundation for reasoning, learning and adaptation. Instead of one system doing everything, many capabilities would build on the same base. The expected benefits are greater adaptability, reasoning that is easier to evaluate, and capabilities that are easier to improve and to combine. Dyssonance's stated mission is to rethink the foundations of AI, so that progress is measured by systems that continually question, learn, adapt and make better sense of the world, rather than by bigger models, faster answers or more powerful memory.

Prendki ends with a line that captures the whole argument: the future of AI isn't about remembering more, it's about understanding better.